Efficient privacy-enhanced familiarity-based recommender system

Recommender systems can help users to find interesting content, often based on similarity with other users. However, studies have shown that in some cases familiarity gives comparable results to similarity. Using familiarity has the added bonus of increasing privacy between users and utilizing a smaller dataset. In this paper, we propose an efficient privacy-enhanced recommender system that is based on familiarity. It is built on top of any given social network (without changing its behaviour) that already has information about the social relations between users. Using secure multi-party computation techniques and somewhat homomorphic encryption the privacy of the users can be ensured, assuming honest-but-curious participants. Two different solutions are given, one where all users are online, and another where most users are offline. Initial results on a prototype and a dataset of 50 familiar users and 1000 items show a recommendation time of four minutes for the solution with online users and of five minutes for the solution with offline users.